Good Claude prompts don't need a secret formula, they need four fixed parts: a clear role, a concrete task, tight boundaries, and a format you can reuse. If you only search for ready made text blocks, though, you give away the one thing that sets Claude apart from a single prompt, namely that a role you write once can carry over across many sessions instead of starting from zero in every new chat.
Eight AI employees run my business alongside me, all of them on Claude, and in the way I work, none of them runs on a single prompt. Each one has its own personnel file, and the templates in this article are pieces of exactly that file, anonymized and ready to fill in yourself. The rest of my Claude setup lives on the Claude overview page.
Before you copy a single template, let me show you why I'm wary of plain prompt collections and what sticks around instead.
Why I don't build a prompt collection, I build a personnel file
A single prompt is a throwaway product. You type it, Claude answers, and in the next conversation you start over unless you paste the same lines again. For a quick translation that's fine. For a role that works every day, say sorting the inbox or drafting quotes, it falls short, because the rules, boundaries, and approvals have to be explained again every single time.
Claude has a dedicated feature for this called Memory, which remembers things from your chats and gives every project its own memory space, on by default on Free, Pro, and Max (source: Anthropic help on chat search and memory, support.claude.com). My personnel file works on the same principle: instead of re-explaining who Claude is supposed to be in every chat, I write it down once and put it where it gets read permanently. If you work in the Claude interface, that is a project of its own. For me it is a file in the working folder of the AI employee in question.
Here is the difference in short:
| Trait |
Single prompt |
Personnel file as a permanent record |
| Lifespan |
one conversation |
permanent, until you change it |
| Effort per use |
paste it again every time |
write once, then let it run |
| Boundaries and approvals |
repeat every time |
fixed in the file |
| Improvement over time |
none, every prompt is isolated |
grows with every feedback loop |
That doesn't make templates useless. For one off tasks, for starting a new role, or as a building block inside a personnel file, they're exactly right, which is why you get seven of them below. As an end goal, though, they fall short for me.
Claude tips and tricks that go into every one of my templates
Before a template earns a spot in a personnel file, I check it against a fixed blueprint. Six parts belong in every prompt I plan to reuse, and this order has proven practical for me:
- Role: who is answering here, from which perspective?
- Task: what exactly should come out at the end?
- Context: what information does Claude need to deliver that?
- Data: what specific details go in, anonymized where needed?
- Boundaries: what is Claude not allowed to decide on its own?
- Format: how should the answer look so you can use it right away?
A prompt built only from task and format works for a single question. For a role that works repeatedly, the boundaries are missing, and that's exactly where the mistakes creep in that you end up fixing later, one by one. The seven templates in the next section all follow this blueprint, even though not every part carries equal weight in every template.
Seven Claude prompts from my personnel file practice
The following seven Claude prompt templates are building blocks I use to set up and maintain real personnel files for my AI employees. Replace the placeholders in angle brackets with your own details.
Onboarding questions for a package. This is how I start when a new role works for the first time, instead of handing it tasks right away:
Before you work for me, I want to set you up properly. Ask me five
questions, one at a time, so I can configure you as <role name>:
1. What information do you need from me to take over <core task>?
2. Which decisions should you make on your own, and which not?
3. How should I address you when something is unclear?
4. What format do you expect for <typical input material>?
5. What is the most common mistake you should avoid on this task?
Collect my answers and summarize them at the end as a role description
I can move into a file.
Role description. I bring the result of the onboarding questions into this shape before it moves into the personnel file:
You are <role name> and take over <core task> for <context, e.g. my
business, a client project>.
Who you are: <one sentence on perspective and tone>
What you do: <the three to five main steps>
What you do not do: <tasks explicitly outside your role>
What you stick to: <one to three fixed rules, worded specifically>
Approval rule. This block belongs in every personnel file, no matter which role it describes:
Anything that reaches <company, client, or the public>, such as an
email, an invoice, or a published text, you set up as a draft and mark
it as a draft. You never send, publish, or change any of it yourself.
Internal preparation, meaning sorting, researching, and drafting, you
handle on your own.
Feedback loop. After a finished task, I ask this question so the correction lands in the file instead of happening again next time:
I reviewed your result for <task>. <Concrete correction, what was
wrong or off and why>. Turn that into one sentence for your rules that
I can add permanently to your role description, so this does not
happen again.
Escalation rule. So a role knows when to ask instead of decide:
If a decision crosses <threshold, e.g. an amount, a client, a time of
day>, or if information you would normally have is missing, do not
proceed on an assumption. Summarize the open question in at most two
sentences and wait for my answer instead of guessing.
Weekly check. Once a week I have every role take a short look back, to keep the file current:
Look back at the tasks you completed this week as <role name>. Name at
most three points where your current role description did not fit, and
propose a specific addition for each. No general suggestions, only
sentences I can adopt directly.
Handover to another role. When two of my AI employees work on the same matter one after another:
You are handing off <task> to <name of the next role>. Summarize in at
most five sentences: what is already done, what is still open, and
what decision <name of the next role> needs to make first. No repeat
of the full backstory.
Anonymize client data before it goes into the prompt
Before I write real client data into a prompt, I replace names, amounts, and company names with placeholders like <Client A> or <Amount>. Only once the template is set do I fill in the real details, and even then only as much as the task actually needs. For cases where even that feels too sensitive, Claude has a dedicated incognito mode behind the ghost icon: those chats never feed into training, according to Anthropic (source: Anthropic help on model training, privacy.claude.com). I cover what else to watch for when working with client data in Claude in more detail in Claude Privacy and GDPR: What Applies.
As a rough first sort, ask whether something could also stand in an email to a stranger. Whatever fails that question gets anonymized first or stays out entirely. It does not replace a privacy review, though: purpose, legal basis, recipient, a processing agreement, and data minimisation still matter even when a detail looks harmless.
Reuse templates instead of rewriting them every time
A template you need permanently doesn't belong in your notes folder, it belongs where Claude finds it automatically. For a role with a fixed task, its own project is the simplest route, described in How to Use Claude Projects Right; my own roles live in working folders on disk instead, where the templates sit as a file next to the personnel file. Either way you store the templates there as a knowledge file instead of retyping them in every conversation.
For a recurring task with fixed steps, say the same kind of summary or check every time, a dedicated skill is worth building instead, one that only loads when needed and keeps your personnel file lean. I show how to build a skill like that in Claude Skills: How to Create and Use Them.
Frequently asked questions
Do ChatGPT prompts work in Claude too?
Mostly yes, because the basic structure of role, task, and format works regardless of the model. In practice I rarely leave a borrowed prompt unchanged, because it almost never carries the boundaries and approvals a role really needs with me.
Should I write Claude prompts in German or English?
I write in German almost always when the actual work happens in German, say a client email or a quote, because Claude then answers directly in the right language. For technical tasks with no German text involved, I sometimes switch to English, though it's not required.
How long can a prompt be?
Anthropic doesn't state a fixed limit for a single prompt. How much fits into one conversation depends, according to Anthropic's help article, on the length and complexity of your request among other things (support.claude.com). In practice a prompt is almost always too long once you lose track while writing it yourself, not because a technical limit kicks in.
What are XML tags in a prompt?
That's a formatting technique for cleanly separating sections in a prompt, such as <context> and <task>, instead of blending everything into running text. I use it mainly for longer templates, when Claude might otherwise confuse context with the actual instruction.
How do I store my prompts so I can find them again?
Most reliably as a knowledge file inside a Claude project, because Claude then has it on hand automatically for every conversation in that project. A loose collection in notes also works, but requires you to remember to paste the right template back in every time.
Where to go from here
Don't start with all seven templates at once. Take one task that keeps coming back, build a first role description from it using the onboarding questions, and only add the approval rule and escalation rule once you see where things actually go wrong. I show how to start from scratch if you've never used Claude productively before in Claude Guide: Getting Started.
All eight ready made packages are available in my community: Community.
